Anthropic is adding machine-readable watermarks to text generated by new Claude models, alongside signed metadata for files. The system applies worldwide to supported models introduced after August 2, as Europe’s new AI-transparency requirements take effect.
The striking detail is that provenance can follow content through copying and pasting. Even proofreading, translating, or formatting human-written material may leave an AI signature. Meanwhile, substantial rewriting or mixing text from several sources can weaken detection.
Why it matters
AI disclosure is becoming part of the content-production infrastructure—not merely a label added before publication. Publishers and businesses will need clearer policies distinguishing AI assistance from AI authorship.
Watermarks may improve accountability, but they could also classify substantially human-created work as AI-generated. Product teams should therefore treat provenance as one signal within a wider editorial record, not as a complete verdict about authorship.
This is part of a broader shift toward layered provenance systems. Metadata can carry detailed context, while embedded watermarks aim to survive transformations that strip metadata. Neither approach is foolproof on its own.